US 8086937 B2
An efficient method to apply an erasure encoding and decoding scheme across dispersed data stores that receive constant updates. A data store is a persistent memory for storing a data block. Such data stores include, without limitation, a group of disks, a group of disk arrays, or the like. An encoding process applies a sequencing method to assign a sequence number to each data and checksum block as they are modified and updated onto their data stores. The method preferably uses the sequence number to identify data set consistency. The sequencing method allows for self-healing of each individual data store, and it maintains data consistency and correctness within a data block and among a group of data blocks. The inventive technique can be applied on many forms of distributed persistent data stores to provide failure resiliency and to maintain data consistency and correctness.
1. An apparatus, comprising:
one or more processors;
a set of data stores in which data blocks and their associated recovery blocks are stored using an erasure encoding and decoding scheme, the set of data stores interoperably coupled to the one or more processors, the one or more processors programmed with code;
the code executable in the one or more processors during erasure encoding to build and maintain a data structure in a memory, the data structure storing information associated with the data blocks and their associated recovery blocks that identify a sequence of changes to the data blocks, and the code executable in the one or more processors following a given failure event to use the information in the data structure to facilitate a recovery operation, the one or more processors programmed with the executable code to perform at least one of steps (a)-(c):
(a) storing a sequence number in a data structure having positions corresponding to the data blocks and their associated recovery blocks;
(b) as a given data block is changed, (i) incrementing the sequence number, (ii) associating the incremental sequence number with the given data block, and (iii) associating the incremental sequence number with recovery blocks in the data structure; and
(c) upon a given failure event, using the sequence numbers in the data structure to recover a data set associated with the sequence numbers in the data structure.
2. The apparatus as described in
3. The apparatus as described in
4. The apparatus as described in
5. The apparatus as described in
6. The apparatus as described in
7. The apparatus as described in
identifying a set of uncorrupted blocks;
determining whether the set of uncorrupted blocks are consistent;
if the set of uncorrupted blocks are consistent, recovering the data set using information in the set of uncorrupted blocks;
if the set of uncorrupted blocks are not consistent, locating one or more recoverable subgroups of the set of uncorrupted blocks; and
recovering the data set using information in the one or more recoverable subgroups of the uncorrupted blocks.
This application is a continuation of U.S. patent application Ser. No. 11/198,061, filed Aug. 5, 2005 now U.S. Pat. No. 7,681,104, titled METHOD FOR ERASURE CODING DATA ACROSS A PLURALITY OF DATA STORES IN A NETWORK. U.S. patent application Ser. No. 11/198,061 claims priority from U.S. Provisional Patent Application No. 60/599,962, filed Aug. 9, 2004. U.S. patent application Ser. No. 11/198,061, U.S. Provisional Patent Application No. 60/599,962, and U.S. patent application Ser. No. 11/198,062, filed Aug. 5, 2005 are incorporated herein by reference.
1. Field of the Invention
The present invention relates to a novel consistency management method that applies redundancy encoding and decoding of data blocks across a plurality of interconnected data stores in a network.
2. Related Art
Erasure coding is an error correction encoding and decoding scheme. Erasure coding is applied to a set of data stores to generate one or more redundant data blocks to protect against erasure of the actual data. A data store refers to a persistent memory for a given data block. In the event of data loss, part of the remaining original data blocks and part of the redundant data blocks can be used to recover the entire original data set. In the event of a device failure (e.g., a data store failure), and when a replacement device is available, the recovered data blocks can be used to reconstruct a latest consistent state of the failed device for distribution to the replacement device.
There are many different types of erasure or error correction coding known in the art. These include, without limitation, data mirroring, parity coding, and algebraic-based coding. Data mirroring and parity coding generally create one additional data block from a number N of original data blocks. This type of coding scheme allows a single data set to survive through one failure while still having the capability to reconstruct the original data. Multi-dimensional parity coding may be applied across several data sets to allow for two or more concurrent failures. Such multiple dimensional parity coding supports multiple failures by combining multiple encoded data sets. Thus, for example, in the case of the two-dimensional parity coding, vertical and horizontal data sets individually allow only one failure, whereas the combination of both data sets allows for two failures. Algebraic-based coding schemes, such as a Reed Solomon code, take N data blocks and generate N+M data blocks. This well-known process is illustrated in
In general, when a data failure occurs, this type of algebraic-encoding scheme requires only any random N copies of data blocks from the N+M number of data blocks to reconstruct the lost data. Thus, algebraic encoding supports up to M concurrent failures in a single data set. To apply algebraic-based coding, when an encoding process detects a data change from one data store, it must generate and update all M redundant data blocks. In other words, it is required that the process have the capability to ensure all M redundant data blocks are completely updated. Because the process may fail during the update (during which other failures may also occur simultaneously), there needs to be a self-healing technique to recover the data from the failure(s).
When applying multiple redundancy erasure coding (such as algebraic-based coding) to data blocks in a set of data stores, one also needs to consider the consistency of the entire data set as well as the correctness of the data blocks. A set of data blocks is considered to be consistent if all the redundant blocks are generated from all the original data blocks. For example, in
While multiple redundancy erasure coding could increase data reliability, it has not been possible to apply it to persistent data stores that are being constantly updated. To address this deficiency, there needs to be an efficient and simple consistency management method in an encoding process to apply the erasure coding. Such a consistency management method would allow data stores to self-heal from failures, and it would ensure data consistency and correctness among all the data blocks.
The present invention addresses this need in the art.
The invention provides an efficient method to apply an erasure encoding and decoding scheme across multiple dispersed data stores that receive constant updates. A data store is a persistent memory for storing a data block. Such data stores include, without limitation, a group of disks, a group of disk arrays, a distributed storage area network (SAN) with virtual data blocks or virtual volumes, or any other standalone or distributed systems, machines, or devices, that hold content fragments over LAN and WAN. The data blocks may contain, by way of example, raw block data, database records, fragments of files, or any fragments of application information. As will be seen, the invention allows for self-healing of each individual data store, and it maintains data consistency and correctness within a data block and among a group of data blocks. The inventive technique can be applied on many forms of distributed persistent data stores to provide failure resiliency and to maintain data consistency and correctness.
A more specific aspect of the invention is an encoding process that applies a sequencing method to assign a sequence number to each data and checksum block as they are modified and updated onto their data stores. The method preferably uses the sequence number to identify data set consistency.
Another more specific aspect of the invention is a recovery process that uses a sequencing method to identify the state of a set of data blocks, and that operates to fix inconsistent data to bring a set of data stores to a consistent state.
Another more specific aspect of the invention is a reconstruction process that recovers lost data from data store failure, and that places the recovered data into replacement data stores.
According to the one or more described embodiments and variations thereon, the invention provides a simple and efficient method and apparatus for providing failure resiliency in distributed and persistent data stores, for maintaining the consistency and correctness of the data in the data stores, and for ensuring the recoverability of the data upon multiple concurrent failures.
A specific embodiment of the invention provides an improved mechanism to provide failure resiliency in a disk array by applying erasure coding to generate multiple redundant data blocks residing in the data stores in an array of disks. This embodiment allows for the data blocks in the disk array to survive through and recover from multiple concurrent device and processor failures. The technique described below allows for self-healing of the disk array to a consistent and correct state. It enables data to recover from multiple device and processing failures, and to enable the data stores to be reconstructed.
In an illustrative embodiment, the method is defined by a set of one or more unordered steps including: (1) identifying N number of original data blocks residing in a data store, (2) locating M number of additional data stores for storing redundant data blocks (checksum blocks), (3) assigning a sequence number to each of the data and checksum stores, (4) applying a given erasure encoding scheme to generate M checksum blocks from N data blocks, and storing the checksum blocks in their associated data store, and (5) for decoding, applying a given erasure decoding scheme to generate all data and checksum blocks from a consistent set of N number of data or checksum blocks. The consistency of a data set is determined using the sequence number assigned to each block. The data and checksum blocks together are sometimes referred to herein as a “data set.”
The foregoing has outlined some of the more pertinent features of the invention. These features should be construed to be merely illustrative. Many other beneficial results can be attained by applying the disclosed invention in a different manner or by modifying the invention as will be described.
The accompanying drawings, which are incorporated herein and form part of the specification, illustrate the present invention and, together with the description, further serve to explain the principles of the invention and to enable a person skilled in the pertinent art to make and use the invention. In the accompanying drawings:
As described above, the present invention relates generally to a method and system for maintaining data correctness and consistency while applying erasure coding to a distributed set of data stores. In particular, the invention provides a highly-efficient sequencing technique for performing erasure encoding and decoding for multiple concurrent data stores, while maintaining data consistency and integrity even during processing and device failures. In general, the purpose of this invention is to provide failure resiliency of the data stores by allowing automated self healing and maintaining data consistency and correctness.
The invention is described with reference to specific architectures and protocols. Those skilled in the art, however, will recognize that the description is for illustration purposes only. One embodiment of the invention provides a method to apply multiple redundancy erasure coding to a disk array, and to enable that array to self-recover to a consistent state under multiple concurrent failures. Another embodiment of the invention provides a method to apply multiple redundancy erasure coding in two dimensions to a group of disk arrays. These embodiments, however, are not limiting, as the principles of the invention may be applied in any standalone or distributed processing or data storage environment.
The representative embodiments are described in more detail below.
A. Multiple Redundancy Coding
As noted above, one form of multiple redundancy coding is an algebraic-based code, such as Reed Solomon, Rabin's Information Dispersal Algorithm (IDA), or the like. This type of code interprets symbols larger than bits as elements, and uses the elements to calculate parity or checksum symbols. The specific algebraic-based code that may be useful here is beyond the scope of this invention. It is merely assumed that one such code is used for the implementation.
As illustrated in
This type of algebraic-based code is very powerful if used on persistent data stores (such as disks or distributed devices) with persistent memory for constant read and write operations. To apply this type of code on persistent data stores that update constantly, however, one must have a technique that can manage random processor or device failures during the encoding and decoding process. Such a technique must have the capability of recovering from a failure during encoding, it must maintain data correctness and data store consistency, and it must provide automated self-healing of the data stores. It must also work appropriately with the encoding and decoding scheme to ensure failure resiliency to the persistent data stores. The present invention solves these and other problems associated with the known art, as will now be described
B. Multiple Redundancy Coding on a Disk Array
B.1 Disk Array Configuration
An administrator or other user may configure the size of a data set to be the same as the number of disks in an array. Preferably, and depending on the coding scheme involved, the number of checksum blocks is smaller then the size of the data set. The higher the number of checksum blocks, the more concurrent failures the disk array can sustain; therefore, the higher the reliability of the array. A simple method of data set and checksum block distribution is illustrated in
As noted above,
Once the configuration is done, any convenient erasure coding scheme (e.g., a Reed Solomon scheme) can be applied as data is inserted into the stores. In particular, when data is input to a data store, an encoding process is triggered. During encoding, an erasure encoding process of the present invention uses the configuration information generated by process illustrated in
B.2 Data Set Consistency
By way of background, a data set is said to be “consistent” if all the checksum blocks are encoded using all the actual data blocks. Suppose for example:
A “consistent” data set is then defined as follows:
B.3 Data Set Partition
A data set is “partitioned” if some subsets of the checksum blocks are encoded with some subsets of the actual data blocks. A combination of the consistent checksum and actual data block subsets then forms a data set “partition” as follows:
When a data set is partitioned, each one of the partitions is consistent, but the data set as a whole is said to be inconsistent. An inconsistent data set is illustrated in
B.4 Data Correctness
A “correct” data block is a data block that is completely updated during a write operation. Thus, for example, assume a data block is at state 1 initially when an update (to modify the data block to state 2) occurs. If the update only modifies a portion of the block (e.g., the data block contains some state 1 information and some state 2 information), the data block is said to be incorrect.
B.5 Erasure Encoding Using Sequencing
Assume C is a counter for a data set, N is the number of actual data blocks in a data set, and M is the number of the checksum blocks in a data set. Further, assume there is one registry for each data store for the actual data or checksum blocks and in which a sequence number is recorded.
When an actual data block d is changed and an update to its data store is needed, the erasure encoding process in
Assume a given data set has five (5) data blocks and two (2) check blocks. In the illustrative embodiment, this requires a total of seven (7) registries. As used herein, a “registry” is any accessible portion of memory or data storage. One or more registries generally comprise a data structure (or data array, linked list, or the like), with a given position in the data structure corresponding to a given registry. At time t0, assume the counter C of the data set is initialized to a given value (e.g., zero (0)) and the registries are initialized to the given value as follows.
At time t1, data block number two (2) changes. After executing the encoding process 800 as described above, the registries are now as follows:
Time t1: D2 changed
After that, assume there are the following sequences of changes:
Time t2: D3 changed
In such case, there are now two data set partitions:
In this data set configuration, five (5) blocks (any of the actual data and checksum blocks) are required to reconstruct the entire seven (7) element data set. To bring the entire seven element set to a consistent state, either partition 1 or partition 2 can be used.
Assume that C2 becomes accessible later and no data is corrupted. The method either can roll forward the entire data set to state of t5 by using partition one (1), or it can roll backward the data set to the state t4 by using partition two (2).
Failure Case 2:
At time t5, assume D4 changed, C1 is updated, but that C2 fails to be updated due to the C2 device failure and the device is to be replaced. The registry values are then as follows:
Time t5: D4 changed
In this case, there are now two consistent data set partitions:
Because in this example a minimum of five (5) elements is required to recover the entire data set (due to the encoding scheme used), partition 2 is unacceptable. Thus, the only recovery choice is to roll forward to t5 using partition one (1). The checksum block C2 in this case cannot be recovered, e.g., until a replacement device is ready. During this time, the data set can continue to be modified and updated as long as there is a consistency management method to identify the consistency and recover the data at any time in the future.
Failure Case 3:
At time t5, assume D4 changed, and both C1 and C2 devices fail and replaced. The registry values are now as follows:
Time t5: D4 changed
In this case, there is only one consistent data set partition:
When replacement data stores are inserted, C1 and C2 can be reconstructed to state of t5.
Although in the above examples the encoding method uses an increment of one (1) for the sequence number, this is not a requirement or a limitation of the technique. For example, the counter can be increment by a given amount, e.g., −1, 2, or any number. Moreover, the sequence number itself can be a given function of the counter.
Generalizing, as can be seen, the above-identified sequencing scheme involves several basic steps: (a) initializing a counter; (b) storing sequence numbers in a data structure (such as the registry table illustrated above) having N+M positions corresponding to the data blocks and their associated recovery blocks; (c) as a given data block is changed, (i) incrementing the counter by a given value (e.g., 1) (ii) assigning the value of the counter to the sequence number at a position associated with the given data block, and (iii) assigning the value of the counter to the sequence number at each position in the data structure associated with a recovery block; and (d) repeating step (c) as one or more data blocks are changed. The resulting data structure is then useful to facilitate a recovery process upon the occurrence of a failure event.
B.6 Correctness Determination
The above described encoding method does not have an indicator to determine if a data block is modified completely. To determine if a data block is correct, an indicator, such as a flag or another sequence number register for each data block, can be used. One embodiment is to have the sequence number entered in the header and trailer of the data and checksum block.
B.7 Recovery Process Using Sequence Number
When a device fails or when a software or hardware error occurs, some data blocks may be corrupted. When a failure occurs during the encoding process, the target data set may become corrupted and partitioned (i.e., inconsistent). Thus, according to the present invention, a recovery process is used to recover the data set to a consistent state. The recovery process typically is configured either to roll the data set forward to a most recent state, or to roll the data set backward to the state prior to the failure.
If there are enough good blocks to recover the data set, the routine continues at step 950 to perform a consistency check. This step is illustrated in
Referring now back to
If (as a result of the processing in
Processing then returns to the main loop in
The routine shown in
To recover to an older state (roll backward), a partition of a lower sequence number can be selected in step 982, with the rest of the recovery process being as otherwise described.
B.8 Reconstruction Process Using Sequence Number
When a data store is bad, the data block it holds is lost. When a new data store is inserted as a replacement, the data block in the bad data store needs to be reconstructed into the replacement data store.
B.9 The Sequence Number Registry and Update Completion Flag
A method to apply the sequencing technique for erasure coding to disks is to create disk blocks (as data stores) to include a header and a trailer. The header and trailer are then used to wrap around the actual data or checksum block. This data structure 1100 is indicated in
C. Multiple Redundancy Coding on Disk Arrays
The present invention provides numerous advantages. In particular, when erasure coding is applied to distributed and persistent data stores that support multiple concurrent failures, consistency management according to the present invention ensures recoverability. This is especially important when the data stores are scattered among many different devices over local or wide area networks. The invention automatically maintains and ensures data set consistency and correctness. The method enables self-recovery of a set of data blocks to a consistent and correct state after encountering one or more failures. A by-product of the invention is a set of failure-resilient, self-managed data stores.
Although not meant to be limiting, the present invention may be implemented in an enterprise such as illustrated in
While the present invention has been described in the context of a method or process, the present invention also relates to apparatus for performing the operations herein. As described above, this apparatus may be specially constructed for the required purposes, or it may comprise a general purpose computer selectively activated or reconfigured by a computer program stored in the computer. Such a computer program may be stored in a computer readable storage medium, such as, but is not limited to, any type of disk including optical disks, CD-ROMs, and magnetic-optical disks, read-only memories (ROMs), random access memories (RAMs), magnetic or optical cards, or any type of media suitable for storing electronic instructions, and each coupled to a computer system bus.
While the above written description also describes a particular order of operations performed by certain embodiments of the invention, it should be understood that such order is exemplary, as alternative embodiments may perform the operations in a different order, combine certain operations, overlap certain operations, or the like. References in the specification to a given embodiment indicate that the embodiment described may include a particular feature, structure, or characteristic, but every embodiment may not necessarily include the particular feature, structure, or characteristic.
Having described our invention, what we now claim is as follows.
Citas de patentes